HIPAA and AI chatbots for clinics.

The short answer

"HIPAA compliant" is not a label anyone can buy. HHS does not certify or endorse products. A chatbot is compliant, or not, because of how a covered clinic sets it up: the agreements behind it, what it asks for, where its transcripts go and who can read them.

So there are two jobs. Get a business associate agreement from every company that touches the conversations. And design the chatbot so the conversations hold as little health information as the job allows. The second job makes the first one smaller.

If you are not sure HIPAA covers your clinic at all, start with the test in HIPAA and your clinic website. This guide is general information, not legal advice.

When a chat becomes health information

Protected health information is identifying information about someone's health, care or payment for care, held by a covered entity or by a business working for one. HHS's view is that identifying information a covered clinic collects through its website is generally protected, even before the person becomes a patient.

Read three opening messages to a dental practice's chatbot:

• "Do you open on Saturdays?" Nobody is identified, and nothing is said about anyone's health.
• "Hi, I'm Dana Ruiz. My crown came off, can I come in tomorrow?" A name and a dental problem: protected.
• "Is it normal to still bleed two days after an extraction? My email is in the form." A health question tied to an email address: protected, and a clinical question the chatbot must not answer.

Visitors type more than you ask for. Someone in pain describes the tooth, the tablets they took and their heart condition in one message. Plan for the chatbot to receive health information even when it never asks for any.

Every company in the chain needs an agreement

A business associate is any company that creates, receives, maintains or transmits protected health information for you, and a subcontractor doing that for one of your business associates counts too. For a website chatbot, the chain usually looks like this:

• The company whose chat widget sits on your site.
• The AI service that reads each message and writes the reply. It is usually a subcontractor of the chatbot company, so it needs to be covered by that company's own agreement with it.
• The servers or database where transcripts are stored.
• The inbox, helpdesk or CRM where handovers and booking requests land.
• Your booking tool, if the chatbot writes requests into it.
• Any text message service the chatbot uses to send confirmations.
• Analytics or session recording built into the chat widget.

Ask the chatbot company for its list of subprocessors and which of them are under a BAA. Encryption does not take a company off the list: HHS treats a service that stores encrypted patient information as a business associate even when it holds no key.

A signed BAA is only the start. The company behind ChatGPT says so in its own help pages: accepting a BAA does not by itself make an application HIPAA compliant.

Which AI plans come with a BAA

The same question applies to the AI tools your team uses behind the chatbot, to summarise a handover or draft a reply. As of October 2026, the vendors' own pages say this:

• ChatGPT: personal plans and ChatGPT Business come without a BAA. Its maker offers one for sales-managed ChatGPT Enterprise and Edu accounts, through ChatGPT for Healthcare for organisations, through ChatGPT for Clinicians for individual verified US clinicians, and for its API on eligible developer accounts.
• Gemini: Google's Workspace BAA covers the Gemini app, apart from Gemini in Chrome, and Gemini in Workspace, once an administrator accepts it in the Admin console. Personal Google accounts are not covered.
• Copilot: Microsoft says Copilot and Copilot Chat, used with work accounts, support HIPAA compliance when properly configured, under its business data protection terms. The web searches Copilot runs are not covered by its BAA.

ChatGPT for Clinicians currently lists physicians, nurse practitioners, physician assistants and pharmacists as eligible. Dentists and physiotherapists are not on that list at the time of writing. Plans and names change, so check the vendor's current page and keep a copy of the signed agreement before any patient information goes in.

Consumer AI plans, and ChatGPT Business, come without a BAA. Patient information must never go into them.

Build it to ask for less

The best protection is a chatbot that does not collect what it does not need. For a booking request that means a name, a way to reply, a few words on the reason and preferred times. It does not need a date of birth, an insurance member number, a list of medicines or a medical history. Those belong in your registration and medical history forms.

When a visitor volunteers more, the chatbot should not ask follow-up questions about it or repeat it back. It moves the conversation to the next step:

• Visitor: "I had a root canal last week and it's throbbing. I'm on blood thinners, can I take ibuprofen?"
• Chatbot: "I can't advise on medicines, but our dentists can. If the pain is severe or you have swelling, please call us now on the number at the top of this page. Otherwise I can ask the team to call you back today. What's your name and best number?"

That reply gives no clinical opinion, puts the urgent route first and collects only what a call-back needs. The question about medicines goes to a dentist through the front desk, with the conversation attached. It is the pattern on our dental clinics page, and the same one works for a physiotherapy clinic, where the questions are about exercises and injuries instead of medicines.

Clinical questions and urgent cases

A clinic's chatbot answers questions about the practice: opening hours, services, prices from your list, how booking works. It does not assess symptoms, suggest medicines or say whether something can wait. Those answers belong to a clinician, and a confident wrong one can do real harm.

Write the urgent wording with your clinicians and keep it short. When a message mentions swelling that makes it hard to breathe or swallow, heavy bleeding or a serious injury, the chatbot gives the emergency instruction first and stops trying to book anything.

In California a second rule applies if AI writes messages about a patient's clinical information. A health facility, clinic or physician's office must show a notice that the message was generated by AI, throughout a chat, plus clear instructions for reaching a person, unless a licensed provider reviews the message first. Scheduling and billing messages are outside the rule. Chatbot disclosure laws has the details.

Transcripts: where they live and who reads them

Decide these points before launch, in writing:

• Where transcripts are stored, and under which BAA.
• Who at the clinic can read them, by name or role. Front desk staff need booking requests. They rarely need every conversation.
• How long transcripts are kept. Chats that never became appointments are the easiest to delete on a schedule.
• What a handover email contains. A link to the conversation in a covered system is safer than the whole transcript pasted into an email.
• How you will answer a patient who asks what you hold about them, or asks you to correct it.
• Who at the chatbot company can open conversations, and when.

Before launch we agree these points with the clinic, along with what the chatbot may collect at all. It is the dull part of the project, and it decides whether the rest is safe.

Keep trackers out of the chat

Chat widgets often come with their own analytics, and some report events to ad platforms. An event called "booking request: implant consultation" sent to an ad platform is a disclosure of health information, however useful it looks in a campaign report.

• Switch off any analytics in the widget that the vendor's BAA does not cover.
• Never send transcript text, form fields or the service a visitor chose to an advertising tool.
• Check what the chat window loads from other companies' servers, with your browser's developer tools.
• Count chats and booking requests inside covered systems, and report totals with no names attached.

The tracking rules for website forms apply here in the same way, and our guide on clinic websites and HIPAA sets them out page by page.

Questions to ask a chatbot company

Put these to any chatbot company before you sign, and get the answers in writing:

• Will you sign a BAA, on the plan we will pay for?
• Which AI service writes the replies, and is it covered by a BAA with you?
• Where are transcripts stored, and for how long by default?
• Can we set our own retention period and delete conversations?
• Who on your side can read our conversations, and is that access logged?
• Is any of our data used to improve your product or anyone else's?
• Which analytics or trackers does the chat window load?
• What happens to our data if we leave?
• How will you tell us about a security incident, and how quickly?
• Can the chatbot be limited to answers we have approved, and does it say when it does not know?

Clinics in Europe: the GDPR version

In Europe the same chat is special category data under the GDPR. Health data may be processed only under an Article 9 exception, such as the patient's explicit consent or the provision of health care under the responsibility of a professional bound by secrecy. The chatbot company is your processor and needs an Article 28 contract that also covers the AI service behind it. Ask where conversations are processed: some services handle data outside the European Economic Area, and your contract should say how those transfers are covered.

Since 2 August 2026 the EU AI Act has also expected a chatbot to tell people they are talking to an AI, at the latest at the first interaction, unless that is obvious. Our guide to the EU AI Act for small businesses covers who carries that duty.

Your staff and AI tools behind the scenes

The chatbot is the visible part. In many clinics the bigger risk is a receptionist who pastes a patient's email into a personal ChatGPT account to draft a kind reply. No BAA covers that account, so the paste can be an impermissible disclosure.

Give staff an approved tool under a BAA, or a firm rule to remove every identifying detail first. Is it safe to put customer data into ChatGPT? sets out what each plan does with the text, and an AI policy for staff gives the one page to write the rule down.

How we set one up for a clinic

Our AI chatbots answer only from information the clinic approves and say so when they do not know. Before launch we test them with the clinic's real questions, including the awkward ones: a patient in pain at night, a question about medicines, someone asking whether they can skip a follow-up. The clinic's team tries the chatbot on their own phones before any patient does, and afterwards the clinic can read every conversation and correct the information behind an answer.

Booking requests go to the front desk, or into the booking tool where its API allows, so a booking chatbot takes requests rather than confirming appointments it cannot see. Whether a clinic needs a chatbot at all is a fair question. If most enquiries need a clinician's eye, a short form and a clear page of answers may serve it better.